Goal & Scoreboard — Stop Overnight Agent Runs Drifting

Long autonomous AI runs fail for one reason far more than any other: the model can't tell whether it's done or whether it's wrong. It generates, assumes success, and errors pile up quietly.

This skill fires before any long, vague, or unsupervised handoff — goal modes, overnight agent jobs, scheduled tasks, client deliverables — and forces three lines into writing:

• The finish line, stated as a STATE rather than a direction
• A scoreboard the model can read by itself — a test that exits 0/1, a tool score, a count that must match, a live page checked for a string (ranked strongest to weakest)
• A retry cap, so it stops and reports instead of spiraling

If a job can't produce those, it doesn't run blind: build the checker first, split the job, or downgrade it to supervised drafts. It explicitly blocks the two classic cheats — the model grading its own homework ("8/10") and invented metrics that sound measurable and aren't.

Includes a TEST-line template for scheduled tasks and how to turn a public scoreboard into a client offer that closes itself ("if it's not 100 on PageSpeed, you don't pay").

Use cases

• Check any long or overnight agent job before you walk away from it
• Diagnose why a previous autonomous run drifted or came back wrong
• Add a real pass/fail TEST line to every scheduled task so none of them fail silently for weeks
• Turn a vague client service into a checkable offer